PNY NVIDIA DGX Spark Grace Blackwell AI Supercomputer

The PNY NVIDIA DGX Spark (NVDGXSPARK-PB) is a compact desktop AI supercomputer powered by the NVIDIA Grace Blackwell architecture, delivering data-center-class AI compute performance in a personal workstation form factor — purpose-built for AI researchers, developers, and enterprises running large language models, generative AI, and inference workloads at the desktop.

Key Features

• NVIDIA Grace Blackwell architecture combining Grace CPU and Blackwell GPU on a single module
• Desktop form factor delivering data-center AI performance without a full server infrastructure
• High-bandwidth NVLink-C2C interconnect between CPU and GPU for maximum memory coherency
• Unified memory architecture enabling large model loading across CPU and GPU memory
• Optimized for NVIDIA AI software stack including CUDA, TensorRT, and NIM microservices

Product Specifications

• Part Number: NVDGXSPARK-PB
• Manufacturer: PNY / NVIDIA
• Architecture: NVIDIA Grace Blackwell
• Form Factor: Desktop AI Supercomputer
• Interconnect: NVLink-C2C (CPU-GPU)
• Memory Architecture: Unified (CPU + GPU coherent)
• Software Stack: NVIDIA CUDA, TensorRT, NeMo, NIM

Processor Specifications

• CPU: NVIDIA Grace (ARM Neoverse V2 cores)
• GPU: NVIDIA Blackwell GPU (GB10)
• GPU Memory: 128GB unified LPDDR5X (shared CPU+GPU)
• NVLink-C2C Bandwidth: 900 GB/s bidirectional
• AI Performance: Up to 1,000 TOPS (INT8)
• FP8 Tensor Performance: Up to 200 AI TFLOPS

Performance Specifications

• System Memory: 128GB unified LPDDR5X
• Storage: High-speed NVMe SSD (configuration dependent)
• Networking: High-speed Ethernet (configuration dependent)
• Power: ~170W TDP (compact desktop thermal envelope)
• OS Support: Linux (Ubuntu, RHEL via NVIDIA DGX OS)

Use Cases

• Local large language model (LLM) development and fine-tuning at the desktop
• Generative AI inference for image, text, and multimodal workloads
• AI research and prototyping without cloud dependency or data center access
• Enterprise edge AI deployment requiring a compact, high-performance compute node

Compatibility

• Runs NVIDIA DGX OS (Ubuntu-based) with full CUDA and AI framework support
• Compatible with NVIDIA NIM microservices for optimized LLM inference deployment
• Works with PyTorch, TensorFlow, JAX, and NVIDIA NeMo AI frameworks
• Supports NVIDIA ConnectX networking for multi-node cluster scaling

Frequently Asked Questions

How does the DGX Spark differ from a standard GPU workstation?

The DGX Spark uses NVLink-C2C to unify CPU and GPU memory into a single coherent pool, enabling models too large for discrete GPU VRAM to run efficiently — a capability not available in standard PCIe-connected GPU workstations.

Can the DGX Spark run 70B+ parameter LLMs locally?

Yes — the 128GB unified memory enables running large models like Llama 70B and similar architectures locally without quantization degradation, which is not feasible on standard GPU workstations with limited VRAM.

Is the DGX Spark intended for inference only or also training?

The DGX Spark excels at inference and fine-tuning. Full pre-training of very large models at scale requires multi-node DGX systems. It is ideal for developers and researchers doing model adaptation and deployment.

The PNY NVIDIA DGX Spark NVDGXSPARK-PB redefines what is possible at the desktop — bringing Grace Blackwell architecture, 128GB unified memory, and NVIDIA's full AI software ecosystem to a compact form factor that puts data-center-class AI development capability directly on the developer's desk.